Schrödinger

Machine learning is a feature layer on a product whose value stands without it. The differentiating assets are simulation engines: free energy perturbation for binding affinity, docking, molecular dynamics and quantum chemistry, built over decades and running on physics rather than on learned parameters. The company states its own position plainly, that machine learning is most useful paired with physics based methods rather than as a substitute for them.

Learned components are real and easy to identify. The enterprise design environment carries a machine learning module for training and deploying predictive models on a customer's own data, the company publishes machine learning method work alongside its physics, and it is now marketing an agentic assistant for materials work that plans and runs multi step workflows spanning literature and data analysis, simulation and machine learning. Those sit on top of the engines. The product is physics with learned components layered over it, and that is a description, not a criticism.

The platform is a design environment for human chemists rather than an autonomous designer, which is the correct posture for a licensed tool and structurally limits how much autonomy is at stake. LiveDesign is explicitly collaborative, presenting physics based and machine learning calculations alongside experimental data so that medicinal chemists prioritize compounds before committing to synthesis. The decision to make a molecule stays with the customer's scientists.

The company publishes no confidence thresholds, prediction reliability bands, or guidance on when a computed result should not be relied upon, and that is the disclosure that would let users calibrate trust in the output.

The methods are published and independently benchmarked across a large peer reviewed literature rather than described in marketing terms, product documentation and a knowledge base are publicly reachable, a Python API is published, and the company runs free and certificated courses that teach the underlying theory including a dedicated course on free energy calculations with FEP+.

Named methodological advances are described specifically rather than generically, for example FEP+ Pose Builder as an integrated pipeline feature. A customer can read how the calculation works, reproduce it, and disagree with it.

The architecture is described and the parties around it are not enumerated. What produces an output is unusually clear: the simulation engines are the company's own and physics based, the machine learning module trains models on the customer's own data inside the customer's environment, and the methods are published, so a buyer knows what is computing an answer. That is a short chain by construction for the licensed software.

The hosted and newer components are where the question sits, and nothing answers it: no hosting provider or subprocessor list is published for the cloud offerings, no model provider is named behind the agentic assistant now being marketed, and nothing states whether a customer's compound structures leave its environment when hosted computation is used.

Evidence exists at both the platform and the asset level, which is rare here. At asset level, the Nimbus originated TYK2 inhibitor zasocitinib (TAK-279), designed using the platform, advanced into Phase III, and the company's own SGR-1505 MALT1 inhibitor received FDA Fast Track designation for relapsed or refractory Waldenstrom macroglobulinemia with initial Phase 1 data reported at the EHA Annual Congress.

At platform level the installed base is the evidence: the software is used across the top 20 pharmaceutical companies by revenue and in more than 70 countries, with published customer case studies including Morphic Therapeutic's alpha4beta7 integrin inhibitor. Caveat for buyers: widely cited figures for cycle time or synthesis reduction come from case studies and vendor materials rather than controlled comparisons, so treat magnitude claims as illustrative.

General privacy handling is published and the questions the platform raises are not answered. The privacy policy (last updated 2 June 2025) reaches the software applications as well as the websites, so it is a service level document, but it governs personal information rather than the research data that matters here: proprietary compound structures, target information and assay results a customer puts into the platform, and the customer specific models trained in the machine learning module.

Nothing published states whether customer research data or trained models are retained, whether anything from them informs the vendor's own methods or its therapeutics programs, or what happens to them at the end of a license, which is a pointed question for a company that runs its own discovery pipeline in parallel with licensing the software. Patient data does not enter the platform; the company's separate clinical trial policy covers participants in its own trials. Safety engineering is not addressed.

A privacy document reaches the service, and no position is taken. The privacy policy (last updated 2 June 2025) covers the company's websites and the software applications it makes available, describing what personal information is collected and how it is handled, which is a substantive service level document rather than a website notice. The company also publishes a policy on clinical trial transparency and the protection of clinical trial participants, covering the trials it sponsors.

No statement of business associate status appears, and no scope position says whether the platform can receive protected health information. For software licensed to research organizations that do not ordinarily send patient data, that one sentence would settle the question. Silence is not a stated position.

Controls are described in a filed document with no attestation against a recognized framework. The company publishes no SOC 2 report, ISO 27001 certificate, security page or trust center, and its policy set covers environment, human rights, procurement, conflicts of interest, health and safety and clinical trial transparency rather than information security. What exists is the cybersecurity disclosure the company files as a listed registrant, describing how it assesses, identifies and manages cybersecurity risk and how that is overseen.

A filing describes a program to investors; it is not an examined artifact a customer can rely on, and it names no framework or assessor. For a vendor whose customers place proprietary compound structures and target information into licensed software and hosted services, that is the gap.

Research software, with no device claim and a scope to match. The platform is licensed as computational chemistry software for research; nothing presents it as a regulated device and no clearance is claimed. The company's regulatory positions belong to its medicines rather than to the software: SGR-1505, the company's MALT1 inhibitor, holds FDA Fast Track designation for relapsed or refractory Waldenström macroglobulinemia and has reported initial Phase 1 data, SGR-3515 is in clinical development, and a TYK2 inhibitor designed on the platform by a partner reached Phase III. Fast Track is a development and review mechanism rather than an approval. No product is approved.

Performance is published by the groups that matter for this product, with a method and a failure analysis, though the company does not present any of it as governance. The domain question here is chemical rather than demographic: whether the methods perform unevenly across target classes and chemical space, and whether the company says where they degrade.

Four things answer it. The company published a peer reviewed assessment of its free energy method's accuracy in a Nature portfolio journal and released the assembled benchmark set openly on a public versioned repository, so others can reproduce and extend it. Its published work reports accuracy varying between targets and chemical series rather than quoting one headline figure. It publishes systematic failure analysis: in the protein binding work it examined every case with an error above 2 kcal per mole, quantified them at roughly 9 percent of the data set, and sorted the causes into sampling failures and force field limitations, naming cases that fall outside the method's domain. And it treats domain of applicability as an explicit topic rather than implying uniform reliability. A large pharmaceutical company has separately published its own benchmark of the method across multiple targets and hundreds of ligands, an evaluation the company did not control.

The product documentation carries no model card, no governance framework and no applicability statement. The material for one already exists in the company's own publications.

A published error rate for the product's headline capability, with its method and its benchmark set. The company published a peer reviewed assessment of its free energy perturbation method's accuracy in a Nature portfolio journal and released the assembled benchmark set, so a prospective licensee can see how far predicted binding affinities sit from measured ones, across which systems, and can rerun the comparison on its own targets. The wider peer reviewed literature on the methods is extensive and includes work by outside groups, and the company publishes product documentation, a programming interface and free courses teaching the underlying theory, including one on the free energy methods themselves, so a customer can learn where the method is reliable rather than being told.

Nothing attaches commercially: no warranty, indemnity or remediation obligation applies to a prediction, and the license disclaims the usual. The machine learning module trains on the customer's own data, where no accuracy figure is published or possible in advance.

No clinical record integration exists: this is research software, and no clinical record system is involved. For a pharma or biotech buyer the systems that matter are its research systems, and here one named integration runs in production: the enterprise design environment is an informatics platform that holds a customer's compound and assay data alongside model predictions, it exposes a published programming interface, and in January 2026 a major pharmaceutical company's model workflows were made available inside it, with that environment named as a priority interface for participating biotech companies. That is a named external system whose workflows run in the product.

What is not published is the rest: no list of electronic lab notebook, laboratory information system or compound registry connectors, and no named customer deployment of such a connection.

Well disclosed, because the product is licensed software rather than a partnership. A public release download portal and documented local installation mean the platform can run on customer controlled infrastructure, so proprietary structures need not leave the organization, and cloud scale execution is available for large virtual screens and ensemble calculations. No explicit data residency terms, tenancy model or regional hosting commitments are published.

The commercial platform has no published price, but the terms around it are easy to find. The end user license agreement and terms of use are published in full alongside a set of corporate policies, the two revenue paths are distinguished plainly (a recurring software license, and discovery collaborations where the company takes milestones and royalties rather than fees), modeling services are offered as a separate paid engagement, and academic terms are set out separately, with a university wide academic site license, a web based route to the free energy tools for academic research that excludes commercial use, and a viewer offered free to academia.

No rate card, tier or unit of charge is published for the commercial software, so the cost cannot be sized before contacting sales; the academic terms describe access rather than price, apart from the free viewer. As a listed company Schrödinger reports software revenue, customer metrics and pipeline spend, which describes its own economics rather than what a licensee pays.

Coverage runs very wide, across life science and materials science alike. Within life science the platform spans small molecule discovery (structure prediction and target enablement, hit discovery, hit to lead and lead optimization, formulation) and mixed modalities including antibody design, peptide discovery, enzyme engineering, bifunctional degraders and TCR modeling.

A parallel materials science line covers polymers, organic electronics, energy storage, thin film processing, catalysis and metals and ceramics. Users span the top 20 pharmaceutical companies, biotech, academia and government laboratories across more than 70 countries.